A statistical correction helps VQE escape measurement plateaus

When fewer measurements become a problem

A variational quantum eigensolver (VQE) estimates a molecule’s energy by repeatedly measuring a quantum circuit, while a classical optimizer adjusts the circuit’s parameters to seek a lower energy. The energy is assembled from many Pauli-operator terms. Terms that can be measured together are grouped into cliques, but each group still needs samples, called shots. Since measurement noise decreases only gradually as more shots are collected, allocating a limited budget well matters.

Variance-preserved shot reduction (VPSR) uses a small pilot sample to estimate how much each group fluctuates, then scales the measurement budget in an attempt to preserve the target energy-estimate variance. The difficulty is that pilot estimates are themselves noisy. The allocation rule is nonlinear in those estimated variances, so applying it to a small sample can introduce finite-sample bias. The authors argue that this can misallocate later measurements, degrade energy gradients, and leave an optimizer stalled even when its current energy is not yet satisfactory.

Correcting the allocation, not the optimizer

Ghozaly, Pekik Nurwantoro, Gagus Ketut Sunnardianto, Yanoar P. Sarwono, and Leong-Chuan Kwek introduce delta-corrected shot reduction (DCSR). They expand the nonlinear VPSR reduction factor around the underlying variances using the multivariate delta method, a statistical approximation for how uncertainty in inputs affects a function’s average. Their formulas add second-, third-, and fourth-order bias corrections, using higher moments of the pilot estimates. The intended change is to the shot-allocation calculation; DCSR does not replace the VQE optimizer or require additional quantum samples beyond the pilot information used by VPSR.

The paper evaluates the method in noisy simulations with Qiskit Aer, using hydrogen (H₂) and lithium hydride (LiH) Hamiltonians. The authors model bit-flip and phase-flip errors during state preparation, gates, and measurements. For H₂, they test moderate and severe shot budgets; for LiH, they vary noise while using a larger budget. They use the Adam optimizer with central finite-difference gradients and report averages across ten independent runs.

Across the reported cases, corrected allocations often let optimization pass plateaus seen with baseline VPSR, and higher correction orders can reduce the cumulative measurements needed to reach the paper’s chosen chemical-accuracy threshold. For example, in one severe H₂ sampling setup, the authors report fourth-order DCSR reaching the target around iteration 90 while VPSR remains stalled until roughly iteration 190. In a LiH case with the higher tested noise, third-order correction reaches the threshold while baseline VPSR and second-order correction do not. These are outcomes of the paper’s simulated setups, not demonstrations on a quantum device.

A correction with a noise limit

More correction is not always better. The authors’ stability analysis finds that with very small pilot samples, the fourth-order term can amplify fluctuations; they advise dropping it below a sample size of about 20 in their analysis. In their LiH noise sweep, the fourth-order option is also excluded at the higher error rate, where third order performs best among the tested choices. At still higher noise, the paper reports that all tested correction orders hit a floor above the target, because the energy-gradient signal is overwhelmed.

The evidence is numerical and limited to two molecular examples, specified noise models, parameter settings, and ten-run averages. The derivation is intended to be independent of a particular molecule or ansatz, but these simulations do not establish performance across all molecules, optimizers, or hardware. The useful lesson is narrower: measurement allocation itself can carry a statistical bias, and analytically correcting it may help a VQE use samples more effectively—provided the pilot data are informative enough and higher-order corrections do not amplify noise. This is a November 2026 arXiv preprint, not a claim of demonstrated hardware advantage.